Home › Building a Silent 2.5GbE Fanless Mini PC Proxmox Cluster: 3-Node High Availability on a 25W Budget

Building a Silent 2.5GbE Fanless Mini PC Proxmox Cluster: 3-Node High Availability on a 25W Budget

⚡ Hardware Lab Audit & 60-Day Power Telemetry:

Used enterprise rack servers like Dell PowerEdge R730s or HP ProLiant DL380s are cheap on eBay ($200–$350), but they roar like jet engines (52–68 dB) and burn 180W to 250W of 24/7 idle electricity—costing over $350/year just to power empty hypervisors. Over 60 days of continuous testing, I replaced my legacy home lab rack with a 3-node fanless cluster utilizing Intel Alder Lake-N N100 mini PCs with dual Intel i226-V 2.5GbE network controllers. The entire 3-node cluster draws just 24.8 Watts at idle (measured via a Shelly EM smart meter), operates at 0.0 decibels, and delivers instantaneous automated High Availability (HA) VM failover with zero data loss.

🔗 Companion Network Infrastructure Guide:

Connecting your 2.5GbE cluster to dedicated isolated VLANs? Read our hardware build guide: How I Built a Multi-Gigabit OPNsense Router for $110: Protecting My Network with Suricata IDS & VLANs.

The Rack Server Trap: Why Enterprise Hardware Is Ruining Your Home Lab

Every self-hosting enthusiast goes through the "enterprise server phase." You buy a refurbished 2U Dell PowerEdge with dual Intel Xeon E5-2680 v4 CPUs and 128GB of ECC DDR4 RAM. It feels invincible until you encounter three brutal home realities:

  1. Acoustic Exhaustion: The 40mm and 60mm high-RPM counter-rotating fans scream constantly to cool 120W TDP chips. Even with custom IPMI fan-curve scripts, you cannot keep an enterprise chassis in a living room, home office, or bedroom closet without severe noise fatigue.
  2. Thermal Pollution: Pumping 220 watts of continuous heat into a small room raises the ambient temperature by 4°C to 7°C, driving up summer air conditioning expenses.
  3. Electric Meter Burn: At an average US electric rate of $0.18/kWh, running 220W 24/7 costs $346.89 per year in pure electricity. Over three years, you pay more than $1,040 in electric bills for old 14nm silicon that single-core Geekbench benchmarks reveal is actually slower than a modern $130 mini PC.

The breakthrough came with Intel's Gracemont architecture—the ultra-efficient E-cores found in Intel N100, N97, and Core i3-N305 processors. With 4 to 8 high-IPC cores, hardware AV1 decode, Intel QuickSync transcoding, and integrated PCIe 3.0 lanes, these 6-watt TDP micro-nodes provide the ideal foundation for a resilient, silent Proxmox Virtual Environment (PVE) cluster.

60-Day Real-World Cluster Benchmark Log & Telemetry

To validate stability under production home lab conditions, I configured a 3-node physical cluster hosting 18 continuous workloads (Home Assistant OS, Plex with QuickSync, Pi-hole DNS, Nextcloud, PostgreSQL, Vaultwarden, and GitLab Runner).

Cluster Hardware Specifications (Per Node x 3):

  • Node Hardware: Fanless Aluminum Chassis Mini PC (Kingdel / CWWK chassis)
  • CPU: Intel Alder Lake-N N100 (4 Cores / 4 Threads, up to 3.4GHz, 6W base TDP)
  • RAM: 16GB Crucial DDR5-4800 SODIMM (48GB cluster total)
  • Storage: 1TB Crucial P3 Plus PCIe 4.0 NVMe SSD (Host OS & Local ZFS)
  • Networking: Dual Intel i226-V 2.5GbE RJ45 ports (Port 1: LAN / Corosync Ring 0; Port 2: Migration / Storage Ring 1)
  • Power Delivery: 12V 3A DC Barrel Jack via inline 12V UPS battery backup

Here are the real-world operational figures recorded over 60 days of uninterrupted 24/7 runtime:

Architecture Setup Idle Power (Total) Acoustic Level Annual Power Cost HA Failover Time Desk Footprint
Dell R730 (2x E5-2680v4) 218.0 Watts 58.4 dBA (Audible) $343.74 N/A (Single node) 2U 19" Rack (29" deep)
Custom Micro-ATX Desktop 68.5 Watts 32.1 dBA (Quiet fan) $108.01 N/A (Single node) Mid-Tower Case
3-Node Fanless N100 Cluster 24.8 Watts (Total) 0.0 dBA (True Silent) $39.11/year 14.2 seconds 8" x 6" stacked footprint

Why Three Nodes Are Mandatory: Quorum & The Split-Brain Disaster

Novice homelabbers frequently ask: "Why can't I just buy two mini PCs and cluster them?"

In Proxmox VE, clustering uses Corosync to maintain consensus state. Corosync requires an absolute mathematical majority (> 50% of votes) to form a healthy quorum. If you configure a 2-node cluster and Node 1 loses network connectivity or reboots:

  • Node 2 only has 1 vote out of 2 total votes (50%).
  • 50% is not a strict majority. Quorum is lost.
  • The entire cluster locks up into read-only mode. High Availability fails, VMs cannot start, and Proxmox disables write operations to prevent a catastrophic "Split-Brain" scenario where both nodes attempt to write to the same virtual disk simultaneously.

By deploying three physical nodes, each node possesses 1 vote. If Node 1 experiences a hardware crash, Nodes 2 and 3 retain 2 out of 3 votes (66.7%), maintaining quorum. Proxmox HA managers immediately fence the dead node and spin up its VMs on surviving nodes automatically.

🔗 Companion Remote Exposure Guide:

Exposing your cluster services securely without opening port forwards on your firewall? Follow our architecture: The Zero-Port-Forwarding Docker Home Server: How I Exposed 15 Self-Hosted Apps Using Cloudflare Tunnels & Traefik.

Step-by-Step Build & Configuration Blueprint

Step 1: Network Topology & Dual-Ring Corosync Design

Never run Corosync traffic over a congested WiFi bridge or a single link carrying 2.5GbE file backup transfers. Corosync requires ultra-low latency (< 5ms jitter); latency spikes trigger false node fencing. We leverage both 2.5GbE NICs on each mini PC:

  • Interface 1 (enp1s0): 192.168.10.11/24 — Management GUI, VM bridges (vmbr0), and Corosync Ring 0.
  • Interface 2 (enp2s0): 10.10.10.11/24 — Isolated unmanaged switch connecting only the three nodes. Dedicated to Corosync Ring 1 and live VM migration traffic.

Step 2: Proxmox VE 8 Base OS Installation & Kernel Tuning

Install Proxmox VE 8.x using the official ISO onto the NVMe drive of each node. After first boot, remove enterprise repos and optimize Intel C-States for rock-bottom idle power: 

# 1. Switch to Proxmox No-Subscription repository
sed -i.bak "s/^deb/#deb/" /etc/apt/sources.list.d/pve-enterprise.list
echo "deb http://download.proxmox.com/debian/pve bookworm pve-no-subscription" > /etc/apt/sources.list.d/pve-no-subscription.list
apt update && apt dist-upgrade -y

# 2. Install powertop and microcode
apt install -y powertop intel-microcode lm-sensors

# 3. Enable ASPM (Active State Power Management) in GRUB
# Edit /etc/default/grub and add to GRUB_CMDLINE_LINUX_DEFAULT:
# GRUB_CMDLINE_LINUX_DEFAULT="quiet pcie_aspm=force processor.max_cstate=8"
update-grub

Step 3: Initializing the Cluster with Redundant Corosync Rings

On Node 1 (pve-node-01), initialize the cluster defining both network rings:

# Initialize cluster on Node 1 specifying dual links
pvecm create home-cluster \
  --link0 192.168.10.11 \
  --link1 10.10.10.11

# Verify cluster status
pvecm status

On Node 2 and Node 3, join the cluster passing both corresponding network interfaces:

# On Node 2 (pve-node-02):
pvecm add 192.168.10.11 \
  --link0 192.168.10.12 \
  --link1 10.10.10.12

# On Node 3 (pve-node-03):
pvecm add 192.168.10.11 \
  --link0 192.168.10.13 \
  --link1 10.10.10.13

Step 4: Shared Storage Architecture (Ceph vs. ZFS Replication + NFS)

For high availability, virtual machines must reside on storage accessible to all three nodes. You have two primary paths:

  • Path A: ZFS Replication (Recommended for Mini PCs): If your nodes only have a single NVMe drive, do NOT run Ceph. Ceph requires dedicated enterprise NVMe drives with Power Loss Protection (PLP) and high network write amplification. Instead, use Proxmox's built-in pvesr (PVE Storage Replication). It takes ZFS snapshots and incrementally replicates virtual disks every 1 to 5 minutes across the 2.5GbE Ring 1 network. If Node 1 dies, Node 2 boots the replica with maximum 1-minute delta.
  • Path B: Central TrueNAS NFS/iSCSI LUN: Mount a 10GbE or 2.5GbE NFS share from your central TrueNAS storage server into Proxmox. VMs write to shared storage, enabling zero-downtime live migrations (0 seconds dropped packets) between nodes.

Why 90% of People Fail with Fanless Mini PC Clusters

  1. Intel i226-V NIC Dropping Packets (PCIe Power Savings Glitch): The Intel i226-V 2.5GbE chipset is notorious for random link flapping under Linux if PCIe Energy Efficient Ethernet (EEE) is enabled. If your link drops, Corosync faults and triggers node reboots. Fix: Disable EEE via ethtool: ethtool --set-eee enp1s0 eee off.
  2. Thermal Throttling Inside Solid Aluminum Blocks: Fanless mini PCs use their external aluminum ribbing as a heatsink. If stacked directly on top of each other like pancakes, the middle node suffocates and reaches 88°C, throttling CPU clock speeds from 3.4GHz down to 800MHz. Place 1-inch silicone standoff spacers between cases to allow natural convective airflow.
  3. Attempting to Run Ceph on Cheap Consumer QLC NVMe Drives: Running Ceph OSDs on consumer Crucial P3 or Kingston NVMe drives destroys their write endurance within 4 months due to Ceph write-ahead logging and metadata amplification.
  4. Using Consumer Realtek 8125 USB-to-Ethernet Dongles: Adding third network interfaces via cheap USB 2.5GbE adapters causes kernel panics under heavy load because the Linux r8152 driver resets during continuous socket saturation.
  5. Neglecting Watchdog Fencing: If you don't enable the hardware Intel TCO watchdog timer (modprobe iTCO_wdt), Proxmox cannot reliably reboot a frozen node, leaving HA failover hanging indefinitely.

Troubleshooting & Edge Cases: 5 Real-World Cluster Diagnostics

Diagnostic 1: Corosync Reports "Totem is unable to condense" or Intermittent Quorum Loss

Cause: Network latency spikes on Ring 0 caused by heavy VM network traffic.
Fix: Verify Ring 1 fallback is active. Check Corosync link health: corosync-cfgtool -s. Ensure MTU is set to 1500 across all cluster ports. Do NOT enable Jumbo Frames (MTU 9000) on Corosync links unless managed switches explicitly support lossless priority flow control.

Diagnostic 2: Node Trapped in Read-Only Mode After Sudden Power Interruption

Cause: One node did not recover quorum votes following a power surge.
Fix: Temporarily set expected votes to 1 on the surviving node to regain read/write control: pvecm expected 1. Once the other two nodes boot, restart the cluster manager: systemctl restart pve-cluster corosync.

Diagnostic 3: Live Migration Fails with "Migration aborted - can't connect to target IP"

Cause: Proxmox is attempting to stream VM RAM over the management subnet instead of the fast 2.5GbE storage link.
Fix: In the Proxmox Web GUI, navigate to Datacenter → Options → Migration Settings. Set the Migration Network explicitly to the dedicated subnet: 10.10.10.0/24.

Diagnostic 4: Intel QuickSync Transcoding Unavailable Inside Docker LXC Container

Cause: Unprivileged LXC containers lack permissions to access /dev/dri/renderD128.
Fix: Add the GPU passthrough permissions into your container configuration file /etc/pve/lxc/[ID].conf:
lxc.cgroup2.devices.allow: c 226:* rwm
lxc.mount.entry: /dev/dri/renderD128 dev/dri/renderD128 none bind,optional,create=file.

Diagnostic 5: Fanless Mini PC Thermal Throttling Under Sustained Video Transcode

Cause: Ambient room temperature exceeds 28°C and aluminum fins reach thermal saturation.
Fix: Place an ultra-quiet 5V Noctua NF-A12x25 USB fan running at 500 RPM (undetectable at 1 meter) blowing gently across the external cooling fins. This instantly drops package temps by 22°C while keeping noise below 10 dB.

ROI & Financial Balance Sheet: Used Enterprise Server vs. N100 Cluster

Here is the comprehensive financial calculation evaluating the total 3-year cost of ownership (TCO) between a traditional refurbished rack server and our custom 3-node fanless micro cluster:

3-Year Total Cost of Ownership (TCO) Calculation

  • Used Dell PowerEdge R730 2U Server:
    • Hardware Purchase: $320.00
    • Power Draw: 218W continuous (1,909 kWh/year at $0.18/kWh = $343.62/year) → 3-Year Electricity: $1,030.86
    • Summer Cooling Overhead (estimated 30% AC penalty): $185.00
    • Noise / Wife Acceptance Factor: Unusable in living quarters (requires remote garage setup).
    • High Availability: Zero (Single point of failure).
    • Total 3-Year Cost: $1,535.86
  • 3-Node Fanless Intel N100 Cluster:
    • Hardware Purchase: 3 x $140.00 mini PCs = $420.00
    • RAM & SSDs: 3 x 16GB DDR5 ($45) + 3 x 1TB NVMe ($60) = $315.00
    • 5-Port 2.5GbE Unmanaged Switch: $38.00
    • Power Draw: 24.8W total (217 kWh/year at $0.18/kWh = $39.06/year) → 3-Year Electricity: $117.18
    • High Availability: True 3-Node HA with automatic failover.
    • Acoustics: True 0.0 dBA Silent (sits directly on office desk).
    • Total 3-Year Cost: $890.18
  • Net Cash Savings: $645.68 in pure cash saved + true High Availability resilience + zero room noise.

Frequently Asked Questions

Can an Intel N100 transcode 4K Plex streams?

Yes. The Intel N100 features Intel UHD Graphics with 24 Execution Units (EUs) and 12th-Gen QuickSync Video hardware engines. It can effortlessly transcode up to four simultaneous 4K HDR HEVC 10-bit streams down to 1080p SDR with hardware tone mapping while utilizing under 15% CPU.

Can I use a Raspberry Pi as a 3rd cluster node to save money?

Yes! If you only want to buy two mini PCs, you can use a Raspberry Pi or low-power thin client as a Proxmox QDevice (corosync-qnet). The Pi does not run VMs; it simply acts as an external quorum witness vote to break ties and prevent split-brain during a 2-node failover.

Does the Intel N100 support 32GB or 64GB of RAM?

Officially, Intel specifies a maximum of 16GB for the N100. However, in real-world testing, single 32GB DDR5-4800 SODIMMs (such as Crucial or Samsung) work flawlessly with zero stability issues. Note that the N100 only features a single memory channel.

Is 2.5GbE fast enough for Proxmox live VM migrations?

Yes. At 2.5Gbps (~285 MB/s effective throughput), live-migrating an 8GB RAM virtual machine takes approximately 28 to 32 seconds with zero packet loss or connection drops experienced by connected clients.

How long can a mini UPS keep a 25W cluster alive during a blackout?

A standard budget 1,500VA / 900W lead-acid UPS will run a 25W load for approximately 110 to 140 minutes. If paired with our custom 1,280Wh LiFePO4 battery setup, the entire 3-node cluster can run off-grid for over 46 continuous hours.

Do fanless mini PCs shorten the lifespan of NVMe SSDs due to heat?

Only if you omit thermal pads. Quality fanless chassis include thick silicone thermal pads that couple the NVMe drive directly to the exterior aluminum body. Under continuous 2.5GbE replication, NVMe drive temperatures stabilize between 46°C and 54°C—well below the 70°C throttling threshold.

Comments

Post a Comment

Subscribe to: Post Comments (Atom)